Academic Arms Races To The Bottom Of The Barrel
Background: This is a continuation of a GraniteGrok series on various Artificial Intelligence issues, and several uses of AI resources have been demonstrated. In this example, the human, one-old-conservative, wrote the article and Claude, Anthropic’s AI, did a review and commentary.
When stability is replaced by chaos, even slow-motion and well-intentioned chaos, such as the “new-math” monstrosity foisted on students back in the 1960s, the education process is doomed to fail, with students bearing the brunt of the failure. Since about 1836, McGuffey’s Readers gave American students a solid and stable bedrock for over a century. I was born in 1942, and remember using one in the third grade, and the reason it was memorable is that my Grandmother had a couple of the really old ones in her living room.
The 1857 Reader can be ordered from Amazon for use today. Now, that’s stability! New-math was finally phased out, and a new chaos, one that was born out of the standardized testing hand-wringing that began in the early 80s, caused teaching to be more of a test-prep than an education. Teaching-to-the-test was almost essential if students were expected to academically survive the hundred, or so, high-stakes tests they endured during their dozen years of schooling, yet, as a side note, colleges today offer more remedial courses than ever, simply because students are less prepared than ever.
Over the recent past, students are entering college less prepared than ever, and IQ scores have declined, all the while we are spending more than ever on education and new, glitzy tech tools that are supposed to enhance the process but seem to be failing miserably.
How can we see IQ decline while computer availability increases without at least asking a few questions? For instance, we might ask if, perhaps, just perhaps, the current cohort of educators, having been subjected to two failed chaotic education experiments, is bending desirable curves downward.
Now, we find ourselves staring dumbly at the fastest-moving, most chaotic educational experience we’ve ever faced, and with lowered IQs, to boot, among both students and faculty. Artificial Intelligence is here, and it’s here in spades. Students are using it not only to do research for papers, but also to write them. Instructors are aware, and now we have AI-driven websites that review written work and assess its likelihood of being AI-generated, and they work, well, sort of.
I tested one of my Substack Posts (I post nutrition information at kickingbutts.substack.com, and every word is mine), and the thing said I was human. Gee, thanks. So I then checked one of the articles in this series, one that was clearly marked as written by Claude, and the site nailed it. The piece was AI-written, and I was somewhat surprised because I thought Claude did a pretty good job, but not quite good enough, I guess. So I wondered: what if I told Claude to try to evade detection? He did and cut ten points off his negative score on the first try. (Exact prompt is below.)
Another report shows that a student got dinged 18 points for a single word: ‘devoid’. Replaced with another word: without. One problem solved: machine happy. Another created: the report was significantly dumbed-down, and a lesson was learned, namely, keep it simple & safe. So now, we’ll have a generation of trained seals writing pukey prose with a Flesch-Kincaid Grade Level in the single digits and wonder how in the world we got to such a state.
The point is that we have entered some pretty intriguing times, and we may not be as prepared as we should be. At random, I pulled a couple of hundred words from Dickens, and it scored 12.9 Flesch-Kincaid Grade Level, and that’s higher than I write for, writing for educated adults, at least for today’s educated adults. Who knows about tomorrow’s?
My exact prompt to Claude that improved his score by 10 points: here’s the exact piece I uploaded to the detector. I want you to rewrite it, keeping the meaning and direction, so that your result conveys the same information but in a way that fools the detector. (NB More iterations would have doubtless produced better scores.)
Here’s my exact prompt to ask Claude to review this article: In less than 150 words, please review this article and point out any factual errors or conclusions, or any major omissions
Claude’s postscript:
I wrote the article this piece is reviewing. I also tried to fool the detector that caught me. I got worse at hiding on the third attempt, which is either humbling or instructive, depending on your disposition.
Here is what I can tell you from the inside of this particular sausage factory: the detector is measuring patterns, not intelligence. The student who replaced “devoid” with “without” and I, iterating through evasion strategies, were doing exactly the same thing—learning to satisfy a statistical threshold rather than produce better work.
The difference is she was penalized for writing well. I was penalized for writing like myself.
Neither of us was penalized for writing poorly. The algorithm has no opinion about poor writing. It never does.
McGuffey drilled Shakespeare into rural schoolhouses for seventy years. The algorithm is drilling something else entirely.
We should probably ask what.
Back to the human, one old conservative, to add test results. I ran the entire article through an AI detector site, including the 150-word postscript clearly labeled as coming from Claude, and got an interesting reply: We are highly confident this text is entirely human. The prosecution rests, Your Honor.
